Native Sentry SDKs like sentry-native and sentry-cocoa have to work with signal handlers to trap processes before they crash. This applies to mobile SDKs as well as the native SDK. As signal handlers are notoriously difficult to work with and the restrictions placed on us are quite limiting we have to partially bend the rules of what is acceptable.
The general risks with signal handlers stem from the interruptive nature:
- they are interrupting already running code which means that the state of the environment around is not safe.
- they need a bit of stack space to operate, yet they might be invoked when we are running out of stack.
- other code might also hang on that signal handler and we likely want to invoke those too.
On POSIX, the signal handlers are thus very restricted in what is acceptable. In theory you can only call async safe functions which unfortunately put a lot of limitations in place. In particular not even allocations are possible, let alone most threading functions.
We want to optimize our outcomes for getting good crash reports and getting crash reports altogether. We do not necessarily optimize for not crashing if we crash after we have managed to record a crash report. What this means is that we consider it acceptable if our crash handling code crashes after it has successfully created a crash report to disk, yet we want to avoid doing so unnecessarily.
Thus the bigger risk is actually creating a locking situation where we prevent the application from properly shutting down but hanging on the crash. This is less of a concern for mobile where the operating system will terminate a hung application, but it's a concern for other situations such as desktop or server setups.
Crash reports should generally be persisted on disk and at a later point be uploaded. This can either happen from an already running thread such as we are attempting on Android or upon restart of the application.
This means that the async handler should optimize for dumping crash data to disk in reducing priority. If for instance we can only collect auxiliary information with the risk of crashing in the crash handler, that one should be extracted and dumped after the stack trace has been created.
The recommendation is for libraries to use custom allocation functions which
under normal circumstances use
malloc but after once have entered a crashing
situation in the signal handler switch over to a bump allocator. This is
necessary as malloc can (and will!) hold a lock when entering into a signal
handler which can cause us to either dead lock or crash.
Another issues is that we cannot generally use locks properly once in signal handlers. The way the native SDK is dealing with this situation is to use a spinlock and to mark when we are in the signal handler. That way generic code that might be wanting to wait for a real lock can either additionally block on the spinlock or silently pass if it's in the same thread as the signal handler.
For writing files the
fwrite family of functions must not be used as they
are not async safe. Instead, the underlying
should be used.